[BUG] When set v2=True in quantize_config: TypeError: NativeProcessor.preprocess() got an unexpected keyword argument 'fail_safe' #1843
Description
Here's my code:
from gptqmodel import GPTQModel, QuantizeConfig, BACKEND
from transformers import AutoTokenizer, AutoProcessor
import json
import base64
from PIL import Image
import io
import os
def format_qwen2_vl_dataset(base64, user, assistant):
return [
{
"role": "user",
"content": [
{
"type": "image",
"image": f"data:image;base64,{base64}"
},
{"type": "text", "text": user}
]
},
{"role": "assistant", "content": assistant}
]
def format_qwen2_dataset(user, assistant):
return [
{
"role": "user",
"content": [
{
"type": "text",
"text": user
}
]
},
{"role": "assistant", "content": assistant}
]
def resize_and_to_base64(image_path, target_size=(448, 896)):
try:
# 打开图片
with Image.open(image_path) as img:
# 调整图片尺寸
resized_img = img.resize(target_size)
# 创建字节流对象，用于存储调整后的图片数据
img_byte_arr = io.BytesIO()
# 保存调整后的图片到字节流（保持原格式）
# 如果需要固定格式，可将format参数改为'JPEG'或'PNG'等
resized_img.save(img_byte_arr, format=img.format)
# 获取字节流中的二进制数据
binary_data = img_byte_arr.getvalue()
# 转换为Base64编码并返回字符串形式
return base64.b64encode(binary_data).decode('utf-8')
except FileNotFoundError:
return f"错误：未找到图片文件 {image_path}"
except Exception as e:
return f"处理图片时出错：{str(e)}"
def prepare_dataset_vlm(format_func, n_sample: int = 20) -> list[list[dict]]:
from datasets import load_dataset, load_from_disk
dataset = load_dataset(
"json", data_files="/nvme/sh/shanghai-2/ltg/EST-VQA_modified.json", split="train"
)
dataset = dataset.shuffle(seed=42)
dataset = dataset.select(range(n_sample))
return [
format_func(resize_and_to_base64(f'/nvme/sh/shanghai-2/ltg/EST_VQA/{sample["image"]}'), sample["conversations"][0]["value"], sample["conversations"][1]["value"])
for sample in dataset
]
def prepare_dataset_llm(format_func, n_sample: int = 20) -> list[list[dict]]:
from datasets import load_dataset, load_from_disk
dataset = load_dataset(
"json", data_files="/nvme/sh/shanghai-2/ltg/distill_r1_110k_modified.json", split="train"
)
dataset = dataset.shuffle(seed=42)
dataset = dataset.select(range(n_sample))
return [
format_func(sample["input"], sample["content"])
for sample in dataset
]
calibration_dataset_vlm = prepare_dataset_vlm(format_qwen2_vl_dataset, n_sample=256)
calibration_dataset_llm = prepare_dataset_llm(format_qwen2_dataset, n_sample=768)
quantize_config = QuantizeConfig(
bits=4, # quantize model to 4-bit
group_size=32, # it is recommended to set the value to 128
desc_act=False,
sym=True,
lm_head=False,
damp_percent=0.01,
damp_auto_increment=0.0025,
static_groups=False,
true_sequential=True,
v2=True,
)
modelpath = "/nvme/sh/shanghai-2/ltg/Qwen2.5-VL-3B-Instruct_0.5"
model = GPTQModel.load(
modelpath,
quantize_config=quantize_config,
trust_remote_code=True
)
calibration_dataset = calibration_dataset_vlm + calibration_dataset_llm
print(calibration_dataset[0])
print(calibration_dataset[-1])
model.quantize(calibration_dataset, batch_size=2, backend=BACKEND.AUTO)
model.save(f"{modelpath}_GPTQ-nomse-v2-4.0.0")When I used v4.0.0 everything worked well. After upgrading to v4.2.5, it gave me this error:
Traceback (most recent call last):
File "/nvme/sh/shanghai-2/ltg/GPTQModel.py", line 117, in <module>
model.quantize(calibration_dataset, batch_size=2, backend=BACKEND.AUTO)
File "/root/miniconda3/lib/python3.10/site-packages/gptqmodel/models/base.py", line 493, in quantize
return module_looper.loop(
File "/root/miniconda3/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
return func(*args, **kwargs)
File "/root/miniconda3/lib/python3.10/site-packages/gptqmodel/looper/module_looper.py", line 301, in loop
processor.preprocess(subset[name], buffered_fwd=buffered_fwd, fail_safe=fail_safe)
TypeError: NativeProcessor.preprocess() got an unexpected keyword argument 'fail_safe'
If I remove v2=True from quantize_config then there's no error.